TCS Research at SemEval-2018 Task 1: Learning Robust Representations using Multi-Attention Architecture

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 291–299 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics TCS Research at SemEval-2018 Task 1: Learning Robust Representations using Multi-Attention Architecture Hardik Meisheri TCS Research New Delhi, India hardik.meisheri@tcs.com Lipika Dey TCS Research New Delhi, India lipika.dey@tcs.com Abstract This paper presents system description of our submission to the SemEval-2018 task-1: Af- fect in tweets for the English language. We combine three different features generated us- ing deep learning models and traditional meth- ods in support vector machines to create a uni- fied ensemble system. A robust representation of a tweet is learned using a multi-attention based architecture which uses a mixture of dif- ferent pre-trained embeddings. In addition, analysis of different features is also presented. Our system ranked 2nd, 5th, and 7th in differ- ent subtasks among 75 teams. 1 Introduction In Natural Language processing, Sentiment anal- ysis refers to the degree of positiveness or nega- tiveness of the information presented in the text. Traditionally sentiment analysis is treated as ei- ther a binary classification task (positive, nega- tive) or a multi-class classification task (very nega- tive, negative, neutral, positive, very positive). Af- fect analysis on the other hand refers to detecting discrete sets of emotions present in the text such as anger, joy, sadness etc (Dalgleish and Power, 2000; Plutchik, 2001). Predicting intensities of these emotions to fine granularity can help us bet- ter understand the sentiment and emotions of the writer. Detecting sentiments or affect from text have a number of useful applications. For example, the degree of disgust or anger expressed in customer complaints or reviews can help us decide the prior- ities of issues to look at, or the joy or optimism ex- pressed in customer feedbacks can be a major fac- tor in deciding the marketing strategy for a com- pany. Sentiment or affect analysis for social media text is a challenging task due to the extensive use of slang, frequent spelling mistakes, innovative and unpredictable use of hashtags and extensive use of emojis and smileys. SemEval-2018 Task 1: Affect in tweets, pro- vides data for 3 languages: English, Arabic, and Spanish. For each language, there are 5 subtasks that are presented, 2 Regression tasks, 2 classifi- cation tasks and 1 Multi-Label task. Further de- tails of tasks are presented in section 3. This task was similar to the WASSA shared task (Moham- mad and Bravo-Marquez, 2017) and dataset pre- sented here is the extension of the data presented for WASSA shared task. In this paper, we present our approach to solv- ing these tasks for English language tweets. We have proposed a system which uses various pre- trained embeddings to handle out-of-vocabulary words and emoji present in the text along with cleaning of raw text. In addition, to create a better representation of the text, we use two sets of em- beddings learnt over two different corpus which results in parallel attention mechanism - one set from the twitter space …

Content truncated to 500 words